Alphabet, Amazon, Meta and Microsoft will collectively spend around $650 billion on AI infrastructure in 2026, according to analysis by Bridgewater Associates as Bijan Alizadeh, founding partner of Cypher Capital explains. That is up 67% on 2025 levels, and it tells you everything about where the real story in AI is heading.
Most of the attention has been on software. Large language models, generative tools, productivity applications. These are the visible outputs, and they are genuinely impressive. But they are only possible because of what sits beneath them.
Think of it this way. Infrastructure is the fridge. OpenAI and Anthropic are the brands inside it. Without the fridge, nothing works.
The supply gap is already here
For those who build and finance physical assets, that analogy lands with particular force. AI infrastructure, such as datacentres, require land, planning permission, grid access and skilled labour. It is, in every meaningful sense, a real asset, and it is one of the most supply-constrained assets on the planet right now.
Data centre occupancy rates are approaching 95% in key markets, according to Goldman Sachs Research. Vacancy in some regions sits below 2%. JLL data shows that 92% of the capacity currently under construction in North America is already pre-committed, through binding lease agreements or owner-occupied development.
Hard drive supply is sold out for 2026. Memory chips are committed through 2027. The systems that power modern AI workloads are, in many markets, already fully allocated.
This supply gap is more than a temporary dislocation. It is a consequence of a decade in which capital followed software returns while the physical foundations of the digital economy were underfunded. Now, we must catch up.
Infrastructure economics are being undervalued
Software businesses have long attracted premium valuations, and the logic is sound. High margins, low costs, fast growth. Infrastructure businesses are the opposite: capital-intensive, slow to build and heavily regulated. For years, investors have priced them accordingly.
That gap is now a mispricing. Software growth depends entirely on the availability of the physical infrastructure needed to run it. That’s the compute, the power, the buildings and everything in between. Investors who grasp that dependency will look at the market very differently.
The hyperscalers already have. Amazon, Alphabet, Meta and Microsoft are cutting back on returning cash to shareholders in order to pour more money into physical infrastructure. These are the companies with the most detailed view of what AI actually requires, and they are voting with their capital.
Goldman Sachs estimates that around $720 billion in grid investment alone may be required through 2030 to meet projected data centre power demand. Energy and data centres will be the defining investment theme of the next decade and we are starting to see this begin in earnest.
Where should we focus capital
Not all AI infrastructure investment is equal. There are three distinct areas where capital is needed, each with its own opportunity and its own set of challenges.
Data centres are the most immediate. Demand is outpacing supply and new facilities cannot be built fast enough. Those with the capital and expertise to deliver at scale will be in a strong position.
Power is the bigger long-term challenge. Data centres are only useful if there is electricity to run them, and securing grid connections and permits takes years. No amount of money can shortcut that timeline, making energy the most overlooked opportunity in the space.
The third area is compute hardware. Access to the specialised chips that power AI has become a strategic resource. Those who can secure and finance that hardware at scale are not just building infrastructure. They are acquiring something that everyone else will eventually need.
The cycle is long
Bridgewater has warned that AI spending has entered a "more dangerous phase," and that caution is not without merit. The scale of investment is enormous, the pace is rapid, and a lot depends on demand continuing to grow.
But infrastructure is a long game, and that is exactly the point. Data centres built today will still be running decades from now. The power systems being planned and financed today will outlast the companies currently building them. The foundations being laid now will support AI applications that have not yet been imagined.
Software moves fast and rewards speed. Conversely, infrastructure rewards patience. The real question for investors is whether the physical systems needed to support that growth can be built in time, and who is backing the people building them.